Improved White Matter Tract Segmentation for Unannotated Dataset

Abstract

White matter (WM) tract segmentation is beneficial to brain research, which provides a valuable tool for analyzing brain development and disease. The introduction of convolutional neural networks (CNNs) has greatly improved the accuracy of WM tract segmentation. However, the training of CNNs usually requires extensive manual annotations of WM tracts, which are often difficult to obtain in practical applications. Therefore, in this study, we explore two methods to realize CNN-based WM tract segmentation when there are no manual annotations of WM tracts for the target dataset and improve the segmentation accuracy. The first method generates registration-based pseudo labels for the target dataset to train the WM tract segmentation network. Specifically, we register images of the publicly available annotated dataset to images of the unlabeled target dataset and improve the binarization strategy by taking advantage of the characteristics of registration and WM tracts to generate the soft labels of WM tracts for target dataset. Moreover, we propose the other method to construct loss weighted matrix for network training using the registration information, which reduces the impact of registration error and further improves the segmentation performance. We evaluated the proposed methods with two dMRI datasets. The results show that the proposed methods are effective in improving the segmentation performance of WM tracts when the manual annotations are unavailable.